We introduce ReAL, a reference-based latent diffusion model for super-resolution that removes the text branch entirely and conditions on a single retrieved reference by injecting its key/value pairs into self-attention, avoiding the semantic drift caused by over-reliance on a text-to-image prior.
Bronze Medal, Best Paper Award at IPIU 2026.

We introduce iRAG, which turns reference selection into a retrieval problem: a 16-bit neural hash finds the nearest reference 64× faster than VGG features, a generative model augments a sparse database, and a variance-based filter removes hallucinated references before training.

We replace the JPEG encoder and decoder with neural operators that stay fully backward-compatible with existing JPEG files, so image quality and chroma preservation improve without changing the deployed pipeline.

The model size an implicit neural representation needs for lossless fitting grows exponentially with bit precision. Decomposing the signal into bit planes turns fitting into per-bit binary decisions, lowering that bound and achieving 16-bit lossless representation and quantization down to 1.58-bit ternary weights.

BurstM replaces deformable-convolution alignment with optical flow for one-to-one correspondence and merges the aligned features in Fourier space, so a single model handles ×2/×3/×4 without per-scale retraining.

NIS extends arbitrary-scale super-resolution to image stitching, estimating Fourier coefficients through neural warping to recover the high-frequency detail that learning-based stitching usually blurs away, and blending color and alignment mismatches in latent space.

ABCD restores low-bit images to arbitrary bit depth with an implicit neural function queried by bit coordinates, removing the contour artifacts that appear when legacy content is shown on high-bit displays.

The angle of linear polarization is defined against a sensor-fixed axis, so it rotates twice as fast as the camera. We bake this spin-2 rotation symmetry into the encoder and inject it into a pretrained backbone, cutting depth error (AbsRel) by 23.8% on mirrored and transparent surfaces.
Samsung Electronics industry collaboration.

Instead of diffusing from noise, we shift only the residual between RGB and polarization along a short trajectory, reaching 4-step inference without distillation, 26× faster (2,252ms → 85ms) at higher angle-image SSIM.
Samsung Electronics industry collaboration.

ToMS re-orients multiplane images to the target camera before generation and merges them per depth plane with neural warping instead of a weighted sum, rendering novel views in 1.3s without the per-scene optimization 3DGS requires.